The IEEE Std 802.1Q specs for credit-based shaper require precise transmit decisions within a 125 microsecond window of time.
Even with the Preempt RT patch or similar enhancements, that isn't very practical as software-only. I doubt that software would conform to the standard's requirements. This is analogous to memory, or CPU. . > -----Original Message----- > From: David Miller [mailto:da...@davemloft.net] > Sent: Thursday, October 5, 2017 1:29 PM > To: levipear...@gmail.com > Cc: j...@resnulli.us; vinicius.go...@intel.com; netdev@vger.kernel.org; > intel-wired-...@lists.osuosl.org; j...@mojatatu.com; > xiyou.wangc...@gmail.com; andre.gue...@intel.com; ivan.bri...@intel.com; > jesus.sanchez-palen...@intel.com; boon.leong....@intel.com; > richardcoch...@gmail.com; hen...@austad.us; Rodney Cummings > <rodney.cummi...@ni.com> > Subject: Re: [next-queue PATCH v4 3/4] net/sched: Introduce Credit Based > Shaper (CBS) qdisc > > From: Levi Pearson <levipear...@gmail.com> > Date: Thu, 5 Oct 2017 12:09:32 -0600 > > > It would be a shame if this were blocked due to a missing software > > implementation. > > Quite the contrary, I think a software implementation is a minimum > requirement for inclusion of this feature. > > Without a software implementation, there is no clear definition of > what is supposed to happen, and no clear way for people to test those > expectations unless they have the specific hardware. > > I completely agree with Jiri. Hardware offload first is _not_ how > we do things in the Linux networking.
